Configuration Tuning for Distributed IoT Message Systems Using Deep Reinforcement Learning: Poster Abstract

Abstract

Distributed messaging systems (DMSs) are often equipped with a large number of configurable parameters that enable users to define application run-time behaviors and information dissemination rules. However, the resulting high-dimensional configuration space makes it difficult for users to determine the best configuration that can maximize application QoS under a variety of operational conditions. This poster introduces a novel, automatic knob tuning framework called DMSConfig. DMSConfig explores the configuration space by interacting with a data-driven environment prediction model(a DMS simulator), which eliminates the prohibitive cost of conducting online interactions with the production environment. DMSConfig employs the deep deterministic policy gradient (DDPG) method and a custom reward mechanism to learn and make configuration decisions based on predicted DMS states and performance. Our initial experimental results, conducted on a single-broker Kafka cluster, show that DMSConfig significantly outperforms the default configuration and has better adaptability to CPU and bandwidth-limited environments. We also confirm that DMSConfig produces fewer violations of latency constraints than three prevalent parameter tuning tools.